TOPOLOGY OPTIMIZATION: FUNDAMENTALS

TOPOLOGY OPTIMIZATION: FUNDAMENTALS

Pierre DUYSINX

LTAS – Automotive Engineering

Academic year 2020-2021

1

LAY-OUT

Introduction

Topology problem formulation: Problem statement

Compliance minimization

Homogenization method vs SIMP based

Filtering techniques

Sensitivity analysis

Solution of optimization problems using structural approximations and dual maximization

Vibration Problems

Applications2

INTRODUCTION

3

What is topology?

4

STRUCTURAL & MULTIDISCIPLINARY OPTIMISATION

TYPES OF VARIABLES

– a/ Sizing

– b/ Shape

– c/ Topology

– (d/ Material)

TYPES OF OPTIMISATION

– structural

– multidisciplinary

structural

aerodynamics,

thermal,

manufacturing

Topology optimization

One generally distinguishes two approaches of topology optimization:

– Topology optimization of naturally discrete structures (e.g. trusses)

– Topology optimization of continuum structures (eventually after FE discretization)

6

Why topology optimization?

CAD approach does not allow topology modifications

A better morphology by topology optimization

(Duysinx, 1996)

Zhang et al. 1993

7

TOPOLOGY PROBLEM FORMULATION

8

TOPOLOGY OPTIMIZATION FORMULATION

Abandon CAD model description based on boundary description

Optimal topology is given by an optimal material distribution problem

Search for the indicator function of the domain occupied by the material

The physical properties write

The problem is intrinsically a binary 0-1 problem ➔ solution is extremely

difficult to solve9

MATERIAL DENSITY FUNCTION

Avoid 0/1 problem and replace by a continuous approximation considering a variable density material running from void (0) to solid (1)

– Homogenization law of mechanical properties a porous material for any volume fraction (density) of materials

– Mathematical interpolation and regularization

SIMP model RAMP

Penalization of intermediate densities to end-up with black and white solution

Efficient solution of optimization problem based on sensitivity analysis and gradient based mathematical programming algorithms

0* EE p=

10

IMPLEMENTION OF MATERIAL DENSITY FUNCTION

Implementation of material density approach is rather easy:

– Fixed mesh

– Design variables are element or nodal densities

– Similar to sizing problem

– SIMP law is easy to code

– Sensitivity of compliance is

cheap

– Use efficient gradient based

optimization algorithms as

MMA

11

A FIRST EXAMPLE: GENESIS OF A STRUCTURE

12E. Lemaire, PhD Thesis, Uliege, 2013

TOPOLOGY OPTIMIZATION AS A

COMPLIANCE MINIMIZATION PROBLEM

13

DESIGN PROBLEM FORMULATION

The fundamental problem of topology optimization deals with the optimal material distribution within a continuum structure subject to a single static loading.

In addition one can assume that

the structure is subject to homogeneous

boundary conditions on Gu.

The principle of virtual work writes

14


A typical topology optimization problem is to find the best subset of the design domain minimizing the volume or alternatively the mass of the structure,

while achieving a given level of functional (mechanical) performance.

Following Kohn (1988), the problem is well posed from a mathematical point of view if the mechanical behaviour is sufficiently smooth. Typically one can consider :

– Compliance (energy norm)

– A certain norm of the displacement over the domain

– A limitation of the maximum stress15


Compliance performance: The mechanical work of the external loads

– Using finite element formulation

Limitation of a given local stress measure ||s(x)|| over a sub-domain W2 excluding some neighborhood of singular points

related to some geometrical properties of the domain (reentrant corners) or some applied loads

– Stress measure ||s(x)|| → Von Mises, Tresca, Tsai-Hill…16


The average displacement (according to a selected norm) over the domain or a subdomain W1 excluding some irregular points

If one considers the quadratic norm and if the finite element discretization is used, one reads

Assuming a lumped approximation of the matrix M, one can find the simplified equivalent quadratic norm of the displacement vector

17


The choice of the compliance is generally the main choice by designers.

– At equilibrium, the compliance is also the strain energy of the structure, so that compliance is the energy norm of the displacements giving rise to a smooth displacement field over the optimized structure.

– One can interpret the compliance as the displacement under the loads. For a single local case, it is the displacement under the load.

18


The choice of the compliance is generally the main choice of designers.

– The sensitivity of compliance is easy to calculate. Being self adjoined, compliance is self adjoined, and it does not require the solution of any additional load case.

– Conversely local stress constraints call for an important amount of additional CPU to compute the local sensitivities.

– One can find analytical results providing the optimal bounds of composites mixture of materials for a given external strain field. The problem is known as the G-closure.

19


Finally the statement of the basic topology problem writes:

Alternatively it is equivalent for a given bounds on the volume and the compliance to solve the minimum compliance subject to volume constraint

20

PROBLEM FORMULATION

For several load cases, average compliance

Or better a worst-case approach

– Where k is load case index, K is the stiffness matrix of FE approximated problem, gk, and qk are the load case and generalized displacement vectors for load case k

– and (x) is the local density and V is the volume 21


Minimize compliance

s.t.

– Given volume

– (bounded perimeter)

– (other constraints)

Maximize eigenfrequencies

s.r.

– Given volume



Minimize the maximum of the local failure criteria

s.t.

– Given volume



TOPOLOGY OPTIMIZATION USING HOMOGENIZATION

VSSIMP BASED TOPOLOGY

OPTIMIZATION

23

TOPOLOGY OPTIMIZATION: FORMULATION

Well-posed ness of problem? Discretised problem is ill-posed

– Mesh-dependent solutions

– Recreate microstructures

– Nonexistence and uniqueness of a solution

Homogenisation Method:

→ Extend the design space to all

porous composites of variable density

Filter method / Perimeter method /

Slope constraints:

→ Restrict the design space by

eliminating chattering designs from the design space

2

4

HOMOGENIZATION METHOD

Select one family of microstructures whose geometry is fully parameterized in terms of a set of design variables [Bendsoe and Kikuchi, 1988]

– G closure : optimal microstructure (full relaxation)

– Suboptimal microstructures (partial relaxation)

Use homogenization theory to compute effectives properties: in terms of microstructural geometrical parameters: Eijk = Eh

ijkl(a,b,…)

Difficult to interpret and fabricate the optimal material distribution as it is

Revival interest with arrival of cellular structures e.g. lattice structures made by additive manufacturing

2

5

POWER LAW MODEL (SIMP)

Simplified model of a microstructured material with a penalisation of intermediate densities [Bendsoe, 1989]

Stiffness properties:

Strength properties:

Modified SIMP should be preferred to avoid singularities

Can be related to actual micro geometries [Bendsoe and Sigmund, 1999]

90% of current topology optimization runs

26

ALTERNATIVE PARAMETRIZATION TO SIMP

Alternatively RAMPparameterization (Stolpe & Svanberg, 2001) enables controlling the slope at zero density

Halpin Tsai (1969)

Polynomial penalization (Zhu, 2009):

27

HOMOGENIZATION METHOD

Investigation of the influence of the selected microstructure upon the optimal topology :

– 1° orthotropic v.s. isotropy

– 2° penalization of intermediate properties28

From anisotropic to isotropic materials

Penalization of intermediate densities

SIMP : PENALIZATION OF INTERMEDIATE DENSITIES

29

SIMP with p=2 SIMP with p=3

SIMP with p=4

POWER LAW MODEL (SIMP)

Prescribing immediately a high penalization may introduce some numerical difficulties:

– Optimization problem becomes difficult to solve because of the sharp variation of material properties close to x=1

– Optimization problem includes a lot of local optima and solution procedure may be trapped in one of these.

To mitigate these problems, one resorts to the so-called continuation procedure in which p is gradually increased from a small initial value till the desired high penalization.

Typically:

– p(0) = 1.6

– p(k+1) := p(k) + Dp after a given number of iterations

or when a convergence criteria is OK30

FILTERING TECHNIQUESAND

MESH INDEPENDENCY STRATEGIES

31

Two numerical difficulties

Checkerboard patterns: numerical instabilities related to the inconsistency between the displacement and density fields.

– Appearance of alternate black-white patterns

– Checkerboard patterns replaces intermediate densities

Mesh dependency: the solution depends on the computing mesh.

– New members appears when refining the mesh

– Number of holes and structural features is modified when changing the mesh.

– Stability (and meaning) of solutions? 32

Checkerboard patterns

Babuska Brezzi conditions of discretization schemes

Checkerboard free numerical schemes

– High order FE elements

– Filtering density field solutions ➔ lower order

density fields

– Perimeter constraint

33


34

FE u: degree 2 / Density : constant

Solution with checkerboardsSIMP with p=2FE u: degree 1 / Density : constant

With perimeter constraint


35

FE u: degree 2

SIMP with p=3. F u: degree 1

Perimeter < 60


36FE u: degree 2

SIMP with p=3. FE u: degree 1

Perimeter < 60

Mesh dependency

Mesh independent solution: insure mesh independent filtering of lower size details

– Low pass filter [Sigmund (1998)]

– Perimeter constraint [Ambrosio & Butazzo (1993)]

37

Mesh independency

38

Perimeter < 60

FE u: degree 1

PERIMETER METHOD

Continuous version of perimeter measure

– With the gradient of the density field and the jump []j of the density across discontinuity surfaces j

The continuous approximation of the modulus of the gradient

39

PERIMETER METHOD

The structural complexity of the structure is controlled with a bound over the perimeter

An efficient numerical strategy has been elaborated to cater with the difficult perimeter constraint (Zhang & Duysinx, 1998)

40

FILTERING MATERIAL DENSITIES

To avoid mesh dependency and numerical instabilities like checkerboards patterns, one approach consists in restricting the design space of solutions by forbidding high frequency variations of the density field.

Basic filtering by Bruns and Tortorelli (2001), proven by Bourdin (2001)

41


Other weighting functions

– Gaussian

– Constant

Density filter is equivalent to solving a Helmotz equation Lazarov et al.

With the following Neuman boundary conditions

42


Historically Ole Sigmund (1994, 1997) introduced a filter of the sensitivities

with

For non uniform meshes, Sigmund proposed to use

The smoothed sensitivities correspond to the sensitivities of a smoothed version of the objective function (as well as the constraints) 43


44


45


46

• Mesh dependent• Checkerboard• Non-Discrete

Solut.

HEAVISIDE FILTER

To obtain 0/1 solutions , Guest et al. (2014) modifies the density filter with a Heaviside function such that if xe>0, the Heaviside gives a physical value of the density equal to ‘1’ and if the xe=0, the Heaviside gives a density ‘0’

47

HEAVISIDE FILTER

To obtain 0/1 solutions , Guest et al. (2014) modifies the density filter with a Heaviside function such that if xe>0, the Heaviside gives a physical value of the density equal to ‘1’ and if the xe=0, the Heaviside gives a density ‘0’

Heaviside smooth approximation

– For b→ 0, the filter gives the original filter

– For b→ infinity, the function reproduces the max operator,

that is the density becomes 1 if there is any element in the neighborhood that is non zero.

48

HEAVISIDE FILTER


– For b→ 0, the filter gives the original filter

– For b→ infinity, the function reproduces the max operator,

that is the density becomes 1 if there is any element in the neighborhood that is non zero. 49

HEAVISIDE FILTER


50

• Mesh dependent• Checkerboard• Non-Discrete Solution• Need of continuation / Large number of iterations (>100)

HEAVISIDE FILTER

Heaviside function can be extended (Wang, Lazarov, Sigmund, 2011) to control minimum and maximum length scale

51

HEAVISIDE FILTER

Heaviside function enables a control of manufacturing tolerant designs ➔ robust design

52

F u

THE THREE FIELD APPROACH

Combining density filtering and Heaviside filter give rise to the so called three field topology optimization scheme proposed by Wang et al. (2011), one uses a design field, a filtered field and a physical field whose relations are defined though the following filter and thresholding processes

– Filtering

– Heaviside

53

SENSITIVITY ANALYSIS

54


Study of the derivatives of the structure under linear static analysis when discretized by finite elements.

The study is carried out for one load case, but it can be easily extended to multiple load cases.

Equilibrium equation of the discretized structure:

– q generalized displacement of the structure

– K stiffness matrix of the structure discretized into F.E.

– g generalized load vector consistent with the F.E. discretization

55


Let x be the vector of design variables in number n.

The differentiation of the equilibrium equation yields the sensitivity of the generalized displacements:

The right-hand side term is called pseudo load vector

Physical interpretation of the pseudo load (Irons): load that is necessary to re-establish the equilibrium when perturbating the design. 56


A central issue is the calculation of the derivatives of the stiffness matrix and of the load vector.

In some cases the structure of the stiffness matrix makes it easy to have the sensitivity of the matrix with respect to the design variable

In topology optimization using SIMP model:

The stiffness matrix

And its derivatives

57

SENSITIVITY OF COMPLIANCE

The compliance is defined as the work of the applied load.

It is equal to the twice the deformation energy

The derivative of the compliance constraint gives:

Introducing the value of the derivatives of the generalized displacements:

58

SENSITIVITY OF COMPLIANCE

The expression of the sensitivity of the compliance writes

Generally the load vector derivative is zero (case of no body load), it comes:

59

NUMERICAL SOLUTION OF TOPOLOGY PROBLEMS

USING GRADIENT BASED MATH PROGRAMMING

60

NUMERICAL SOLUTION OF TOPOLOGY OPTIMIZATION PROBLEMS

Optimal material distribution = very large scale problem

– Large number of design variables: 1 000 → 100 000

– Number of restrictions:

1 → 10 (for stiffness problems)

1 000 → 10 000 (for strength problem with local constraints)

Solution approach based on the sequential programming approach and mathematical programming

– Sequence of convex separable problems based on structural approximations

– Efficient solution of sub problems based on dual maximization

Major reduction of solution time of optimization problem

Generalization of problems that can be solved

61

SEQUENTIAL CONVEX PROGRAMMING APPROACH

Direct solution of the original

optimisation problem which is

generally non-linear, implicit

in the design variables

is replaced by a sequence of optimisation sub-problems

by using approximations of the responses and using powerful

mathematical programming algorithms62

SEQUENTIAL CONVEX PROGRAMMING APPROACH

Two basic concepts:

– Structural approximations replace the implicit problem by an

explicit optimisation sub-problem using convex, separable,

conservative approximations; e.g. CONLIN, MMA

– Solution of the convex sub-problems: efficient solution using dual

methods algorithms or SQP method.

Advantages of SCP:

– Optimised design reached in a reduced number of iterations:

typically 100 F.E. analyses in topology optimization

– Efficiency, robustness, generality, and flexibility, small computation

time

– Large scale problems in terms of number of design constraints and

variables 63

Linear approximation and Sequential Linear Programming

Linear approximation = first order Taylor expansion around x0:

When linear approximation is applied to each function of the problem, one transforms the problem into a sequence of linear programming problems (SLP):

64

SEQUENTIAL LINEAR PROGRAMMING METHOD

The current design point is x(k). Using the first order Taylor expansion of f(x), hj(x), we can get a linear approximation of the NL problem in x(k):

Solving this LP problem, we get a new point in x(k+1) and start again.

65

DOESN’T WORK!

SEQUENTIAL LINEAR PROGRAMMING METHOD

MOVE LIMIT STRATEGY

Introduce a box constraint around the current design point to limit the variation domain of the design variables

Of course take the most restrictive constraints with the side constraints

66

STRUCTURAL APPROXIMATIONS

Convex Linearisation (CONLIN)

Method of Moving Asymptotes (MMA)

67

CONLIN approximation

45 90 180

100

105

110

115

120

125

130

135

140

145Strain Energy

(N/mm)

krX

klX

)(Xg

)(~ Xgr

)(~ Xgl

Approximation of the strain energy in a two plies symmetric laminate subject to shear load and torsion (Bruyneel and Fleury, 2000)

68

MMA approximation

Approximation of the strain energy in a two plies symmetric laminate subject to shear load and torsion (Bruyneel and Fleury, 2000)

45 90 135 180

100

105

110

115

120

125

130

135

140

145

kX

Strain Energy(N/mm)

kL

)(Xg

)(~ Xg

*kX45 90 180

100

105

110

115

120

125

130

135

140

145

kX

kU

Strain Energy

(N/mm)

)(Xg

)(~ Xg

*kX

69

DUAL METHODS

Primal problem

Lagrange function:

If the problem is convex…

70

DUAL METHODS

Dual problem

– with

Solve Lagrangian problem

Lagrangian problem

71

NUMERICAL APPLICATIONSOF COMPLIANCE

MINIMIZATION BASED TOPOLOGY OPTIMIZATION

72

x

y-1000N

-150N

1000N

150N

-5000N

-7500N

Optimization of a maximum stiffness bicycle frame

Load cases

Optimum topology

Nonconventional design

Conventional design

73

Design of a Crash Barrier Pillar (SOLLAC)

74

Design of a Crash Barrier Pillar (SOLLAC)

75

Topology Optimization of a Parasismic Building (DOMECO)

76

3D cantilever beam problem

No perimeter constraint

E=100 N/m², n=0.3

20 x 32 x 4 = 2560 F.E.s

77

3D cantilever beam problem

Perimeter = 1000Perimeter = 1400 78

An industrial application: Airbus engine pylon

Application

– carried out by SAMTECH and ordered by AIRBUS

Engine pylon

= structure fixing engines to the wing

Initial Model– CATIA V5 import → Samcef

Model

– BC’s: through shell and beam FE

– 10 load cases: GUSTS

FBO (Fan blade out)

WUL (Without undercarriage landing)

79

Over 250.000 tetraedral FE

An industrial application: Airbus engine pylon

Target mass: 10%

Additional constraints:

– Engine CoG position

Optimization parameters

– Sensitivity filtering: (Sigmund’s filter)

– Symmetry (left right) condition

– Penalty factor

CONLIN optimizer: special version for topology optimization 80

Sensitivities filtering

Penalty factor from 2 to 4

Airbus engine pylon

81With courtesy by Samtech and Airbus Industries

Airbus engine pylon

82With courtesy by Samtech and Airbus Industries

Sandwich panel optimization

Geometry of the sandwich panel reinforcement problem

Optimal topology

83

Sandwich panel optimization

Geometry of the sandwich panel reinforcement problem

Optimal topology

84

PLATE AND SANDWICH PLATE MODELS

(a) (b)

(c) (d)

USING SIMP MODEL FOR TOPOLOGY OPTIMISATION OF PLATES AND SHELLS

0.0 0.25 0.5 0.75 1.0

is replaced by:

PHYSICAL MEANING OF DENSITY VARIABLE:

PROTOTYPE CAR BODY OPTIMIZATION

Load case 1: bending

– Self weight

– Components (20 kg)

– Pilot (50 kg)

– Roll over load (70 kg on top of roll cage)

Load case 2: torsion + bending = curb impact

– Rear axle clamped

– Right front wheel free supported

– Left front wheel withstanding 3 times the weight of the axle

70 kg

weight x 3= 2700 N(Figures from Happian-Smith)

87

DESIGN OF A URBAN CONCEPT STRUCTURE

Topology optimization of the truss structure

– Target mass of 15 kg

– Minimum compliance

– Mostly determined by load case 2 (torsion)

– SIMP material with p=3

– Left / right symmetry of material distribution

– Filtering

8

8

DESIGN OF AN URBAN CONCEPT STRUCTURE

Convergence history89


Volume = 40%

Volume = 20%

Volume = 60%90


9

1

VIBRATION PROBLEMS

92

NATURAL VIBRATION PROBLEMS

Finite element discretization of the system

Kinetic energy

Strain energy

93


Hamilton principle

Dynamic equation of the system

Free vibrations : assume periodic solutions

Nontrivial solutions are solutions of the eigenvalue problem

94


Eigenvalue problem

Rayleigh ratio

95


Fundamental topology optimization problem of vibrating structures

To avoid mode crossing, it is better to select several eigenvalues and to maximize the minimum of the first NF frequencies

96


However, the fact that both mass and stiffness depends on the density design variables, the trivial solution 0=0 is feasible.

Therefore best eigenfrequency design problem is formulated as a reinforcement problem, i.e. there exist some non design mass or stiffness

97

REINFORCEMENT DESIGN PROBLEM

Topology optimization of sheet of steel:

– Basic sheet: t=1mm

– Reinforcement sheet t=1mm

98

REINFORCEMENT DESIGN PROBLEM

Topology optimization of sheet of steel:

– Basic sheet: t=1mm

– Reinforcement sheet t=1mm

99

SENSITIVITY OF EIGENVALUE PROBLEMS

Eigenvalue problem

– K stiffness matrix, M mass matrix

– q the eigenmode vector

– And w the eigenfrequency

The magnitude of the modes is arbitrary, so they are normalized according to a given matrix W (generally the mass matrix M)

At first let’s consider the simplified approach: we assume that all eigenvalues are distinct and ordered from the smallest to the largest:

100


Let’s differentiate the eigenvalue equation

Differentiating the normalization equation gives

To obtain the derivatives of the eigenvalue l(k), one has to

premultiply the first equation by the eigenmode q(k)

101


Since q(k) is an eigenmode

And one gets

With the scaling factor

We finally obtain the final expression of the sensitivity of the eigen values:

102

DIFFICULTIES IN VIBRATION PROBLEMS

Topology optimization of vibrating structures presents a major difficulty:

– Appearance of dummy eigenmodes, i.e. local modes with zero frequency

Require special strategy

– Modification of material interpolation ➔ modified SIMP or

RAMP

– Filtering the dummy modes

103


Illustration using a numerical example

Material : Steel E=210 Gpa, n=0.3, =7800 kg/m³

T=1e-3 m

Interpolation law: SIMP p=3

Max volume= 80%

Design domain = Support only

104


Illustration using a numerical example

Material : Steel E=210 Gpa, n=0.3, =7800 kg/m³

T=1e-3 m

Interpolation law: SIMP p=3

Max volume= 80%

Design domain = Support only 105


Convergence curve after 10 iterations

106


Convergence broken after 4 iterations when SIMP p=4

107


Low frequency modes are present in low density regions

Modify SIMP to give a lower bound in low density to the ratio E/

– Modified SIMP

108


Ignore / filter eigenmodes which are not significant

– Selection criteria: generalized mass ➔ local character of the

mode

109

Date post:	23-Feb-2022
Category:	Documents
Upload:	others
View:	17 times
Download:	0 times

TOPOLOGY OPTIMIZATION: FUNDAMENTALS

Documents